New in v17.18.0 - ✨Agent Client Protocol in CodeCompanion - Code with Agents

[olimorris]

Hi All,

Very happy to announce that in v17.18.0, you can leverage the newly announced Agent Client Protocol from Zed, in CodeCompanion.

TLDR; The Agent Client Protocol allows for code editors like Neovim to communicate with coding agents such as Gemini CLI and Claude Code (although the latter does not yet support ACP).

You can see it in action below (requires audio on to hear my voiceover):

Agent.Client.Protocol.mp4

---

I recommend reading Zed's blog post on how they created the Agent Client Protocol and utilize it, here.

The team at Zed reached out to me at the back end of July and shared their plans for the protocol and what they were hoping to achieve. Turns out they're huge fans of Neovim and wanted us to be one of the first editors to have support for it.

Huge shout out to @ConradIrwin, @agu-z, @benbrandt and @morgankrey for their support and guidance over the last couple of weeks. Amazing to finally have something like this that's free and open source.

---

Currently only Gemini CLI has support for ACP but expect to change in the coming weeks.

Finally, a link to Using Agents in the docs.

[Davidyz]

That's some exciting updates!

Just a heads up for people using extensions (for example, VectorCode query tools with summarisation and codecompanion-history.nvim): if you use an ACP adapter for chat strategy, it might be breaking some of the functionalities if the extension directly talks to the chat adapter, because the ACP adapters don't share the same API as the OG http adapters. You can just configure the extension to use an HTTP adapter and get away with it, though.

[nshen]

Cool, seems I finally don't have to endure Gemini cli's single-line input anymore, but how do we reference other file ? using /file or @ that file?

[olimorris]

Variables and slash commands 👍

[nshen]

The API login has very limited free quota, it would be perfect if it supported Google login.

[olimorris]

I've just been using an API key and haven't investigated this but feel free to PR it.

We receive the authentication methods from the agent but proceed with just the API Key. Turn on debug logging and you should see everything it returns

[viespejo]

@nshen I've tried all three authentication methods and they all work. Make sure you have the environment variables that the Gemini CLI needs for each authorization method authentication docs

You can use different commands (see config code snippet below) to use the "flash" model when using Gemini API KEY and running out of "pro" requests.

Thanks, @olimorris (and everyone else) for such a fantastic project!

			acp = {
				gemini_cli = function()
					return require("codecompanion.adapters").extend("gemini_cli", {
						commands = {
							flash = {
								"gemini",
								"--experimental-acp",
								"-m",
								"gemini-2.5-flash",
							},
							pro = {
								"gemini",
								"--experimental-acp",
								"-m",
								"gemini-2.5-pro",
							},
						},
						defaults = {
							-- auth_method = "gemini-api-key", -- "oauth-personal" | "gemini-api-key" | "vertex-ai"
							auth_method = "oauth-personal",
							-- auth_method = "vertex-ai",
						},
					})
				end,
			},

[olimorris]

@viespejo thanks so much for testing this out and love the additional commands. I'll get this into the docs later today

[bassamsdata]

auth_method = "oauth-personal",

Ah this is amazing, @viespejo. Thank you.

[bassamsdata]

This is honestly one of the most exciting developments I’ve seen in AI and Neovim, especially if the protocol gains wide adoption. Amazing work!

[nathansobo]

This demo video is really excellent. Thanks so much for collaborating on this with us @olimorris! You're awesome!

[olimorris]

I was trying to get remotely close to your video in the blog. This is where you tell me you did it in 1 take and with no script... 😆

[kylesnowschwartz]

If/when ACP is supported by Anthropic and Claude Code - will I need to port over all my existing CC configuration to use CodeCompanion? i.e. the agents, output-styles, custom commands and etcetera. I'm just hoping that I can make a seamless transition to a better terminal UI - in neovim - without having to re-invent all my workflows and keep them in sync with my team at work that has already coalesced around using Claude Code. Thanks! 🙏

[olimorris]

I expect everything to work with your existing Claude Code configuration. CodeCompanion will just be handling the communication between yourself and Claude Code. You may need to alter the command in the Claude Code ACP adapter to pass in your config via a parameter but that's as complex as I expect it to get.

[BobTB]

Not quite related, but I was thinking how to make it not send the whole context for simple questions, wasting the quota. For example, "Get me a regex for ... "
Some command as /nocontext and the request just sends the question and nothing else, this could be a nice token saver which could be implemented here.

[olimorris]

What does "whole context" mean?

[BobTB]

By “whole context” I mean that when I ask a question, it probably sends all of the conversation history plus the files it thinks are need in the workspace along with my prompt. That’s useful when the question is related to the codebase or earlier discussion, but if I just ask something simple (like “give me a regex for …”), the context isn’t actually needed.

The downside is that sending everything every time consumes more tokens and quota than necessary and most of the time we just forget that and instead of switching to some web chat just ask where we currently are...

[olimorris]

This isn't the case at all. From the docs:

A key difference in working with agents versus LLMs is the matter of state. LLMs, via http adapters, are stateless. This means that CodeCompanion sends the entire message history over with every request. Agents differ in that they are the ones responsible for managing state. As a result, CodeCompanion only sends the latest messages over with every prompt. From a UX perspective however, neither of these have an impact on how it feels to work with CodeCompanion.

Perhaps the agents sends the history with every request and that does count towards token consumption, but that isn't a CodeCompanion problem.

[boydaihungst]

auth_method = "oauth-personal". This will force user to log in every time a new codecompanion chat window is opened? Why doesn't it use cached creds file? $HOME/.gemini/oauth_creds.json

[ntcong]

there's a bug in gemini-cli with cached creds in acp mode: google-gemini/gemini-cli#7549

[boydaihungst]

I see. Thanks for the information.

[bassamsdata]

You can monkey-patch the plugin as a temporary solution until the issue is fixed.

If you use lazy.nvim, add a build step for the plugin.

lazy.nvim build

    build = function()
      local plugin_path = vim.fn.stdpath("data") .. "/lazy/codecompanion.nvim"
      local patch_file = vim.fn.stdpath("config")
        .. "/patches/codecompanion/skip_oauth.patch"
      vim.system(
        { "patch", "-d", plugin_path, "-p1", "-i", patch_file },
        { text = true },
        function(obj)
          vim.schedule(function()
            if obj.code == 0 then
              vim.notify("Patched codecompanion.nvim successfully", vim.log.levels.INFO)
            end
          end)
        end
      )
    end,

Place the file skip_oauth.patch wherever you prefer. Mine is at nvim/lua/patches/codecompanion/skip_oauth.patch.

skip_oauth.patch

diff --git a/lua/codecompanion/acp/init.lua b/lua/codecompanion/acp/init.lua
index d9c8da4b..3e91352e 100644
--- a/lua/codecompanion/acp/init.lua
+++ b/lua/codecompanion/acp/init.lua
@@ -170,13 +170,25 @@ function Connection:connect_and_initialize()
       end
       methodId = methodId or (auth_methods[1] and auth_methods[1].id)

-      if methodId then
+      -- Skip OAuth authentication if credentials already exist
+      local should_skip_auth = false
+      if methodId == "oauth-personal" then
+        local oauth_creds_path = self.adapter_modified.defaults.oauth_credentials_path
+        if oauth_creds_path and vim.uv.fs_stat(oauth_creds_path) then
+          log:debug("[acp::connect] Found existing OAuth credentials at %s, skipping authentication", oauth_creds_path)
+          should_skip_auth = true
+        end
+      end
+
+      if methodId and not should_skip_auth then
         local ok = self:send_rpc_request(METHODS.AUTHENTICATE, { methodId = methodId })
         if not ok then
           log:error("[acp::connect_and_initialize] Failed to authenticate with method %s", methodId)
           return nil
         end
         log:debug("[acp::connect_and_initialize] Authenticated using %s", methodId)
+      elseif should_skip_auth then
+        log:debug("[acp::connect] Skipped authentication for %s (existing credentials found)", methodId)
       else
         log:debug("[acp::connect_and_initialize] No compatible auth method; skipping authenticate")
       end
diff --git a/lua/codecompanion/adapters/acp/gemini_cli.lua b/lua/codecompanion/adapters/acp/gemini_cli.lua
index 2bf341fe..bf7f8e77 100644
--- a/lua/codecompanion/adapters/acp/gemini_cli.lua
+++ b/lua/codecompanion/adapters/acp/gemini_cli.lua
@@ -25,6 +25,7 @@ return {
   },
   defaults = {
     auth_method = "oauth-personal", -- "oauth-personal"|"gemini-api-key"|"vertex-ai"
+    oauth_credentials_path = vim.fs.abspath("~/.gemini/oauth_creds.json"),
     mcpServers = {},
     timeout = 20000, -- 20 seconds
   },

[boydaihungst]

Nice. Thank you 😊

[boydaihungst]

Works perfectly. I only had to run Lazy build codecompanion.nvim.